القائمة
Task-conditioned workspace subgraph — nodes, edges, hierarchy.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
Preview a Rust symbol rename — exact reference set & refactor probe.
Audit a Rust crate's public API.
Enforce Rust crate-edge rules.
Audit Rust attributes and doc-comments.
Rust fn-level call graphs.
Rust complexity hotspots by blast radius.
استنادا إلى تصنيف SOC المهني
| name | rmc-codemap |
| description | Task-conditioned workspace subgraph — nodes, edges, hierarchy. |
| argument-hint | <task-prompt-or-seed-list> |
| allowed-tools | Read, mcp__rust-code-mcp__* |
build_codemap returns a focused subgraph of the workspace hypergraph keyed off either a natural-language task_prompt or explicit seed_qualified_names. One MCP call composes BFS expansion + relevance scoring + hierarchy projection + optional snippet extraction into a single artifact: nodes + edges + filtered module tree + stats.
Use this when "what code interacts with X?" is too coarse for rmc-call-graph's single-fn shape. For workspace-wide duplicate detection, use rmc-semantic-overlaps. For exhaustive one-symbol analysis (uses + imports + bindings), use rmc-symbol-forensics. For refactor planning around the result, hand off to rmc-refactor-plan.
build_hypergraph(directory=<absolute-path>)
If the snapshot is older than the newest .rs file in the workspace, build_codemap returns a diagnostic asking for build_hypergraph(directory=..., force_rebuild=true). The check runs automatically.
task_prompt mode also requires the BM25 / vector indices: run index_codebase(directory=...) once.
| Mode | Seed | When to use |
|---|---|---|
seed_qualified_names=[...] | Explicit list of pub / pub(crate) qualified names | You know the entry point. Fast, deterministic. |
task_prompt="..." | Natural-language query → HybridSearch (BM25 + vector) | Exploratory. Best against code with rich doc comments. |
Pass at least one. Both together: seeds take precedence; prompt ignored.
build_codemap(directory=..., seed_qualified_names=[<crate>::Y])
Defaults: max_nodes=80, depth=3, max_incoming_per_node=8, embedding_policy="no_rerank", format="json", include_snippets=false.
| Param | Default | Cap | Effect |
|---|---|---|---|
max_nodes | 80 | 500 | Total retained nodes. Seeds always survive; only non-seeds prune. |
depth | 3 | 5 | BFS depth from each seed, both directions. |
max_incoming_per_node | 8 | — | Per-node incoming-edge fanout cap (ranked by BM25 hit score). |
Focused: max_nodes=15, depth=1. Sweeping: max_nodes=200, depth=3.
format | Best for |
|---|---|
json (default) | LLM ingestion / dashboard / programmatic consumer |
outline | Terminal / chat reading |
mermaid | Diagram for docs |
all | One call → three views (~3× token cost) |
include_snippets=true adds first 5 lines / 400 bytes of each item's source. Aligned to line boundaries; doc-comment included.
embedding_policy | Cost | Quality |
|---|---|---|
no_rerank (default) | Free | 0.60 × bm25_norm + 0.40 × graph_prox |
cached_only | Free | Adds cosine for items already cached |
compute_missing | ~10 ms / node first call | 0.40 × emb_sim + 0.35 × bm25 + 0.25 × graph_prox |
Cache is shared with rmc-semantic-overlaps — running either warms both.
Codemap.diagnostics is a Vec, never errors. Surface to the user when non-empty.
| Diagnostic | Meaning |
|---|---|
"unresolved seed: X" | Override-seed name doesn't resolve, parent path doesn't either. Typo or wrong crate. |
"unresolved seed: X (parent module resolves; leaf likely private or not indexed)" | Parent module is real; leaf isn't indexed. Likely private. Make pub(crate) if you want it navigable. |
"no search hits resolved to graph items" | Prompt path: all hits dropped. Refine the prompt or use override seeds. |
"{n} search hits dropped: {a} path-norm, {b} line-resolve, {c} kind-filter" | Funnel counters. High line-resolve → prompt too broad; hits landed in import lines or comments. |
"snapshot is older than newest .rs file ..." | Rebuild the snapshot. |
build_codemap(directory=..., seed_qualified_names=[<crate>::X], depth=2)
One seed, both directions, depth 2. Returns the cohort of callers + callees + types X interacts with. ~30-60 nodes typical.
build_codemap(directory=..., task_prompt="how does <subsystem> work",
depth=3, format="outline")
Exploratory. Check diagnostics for drop counts — high line-resolve means refine the prompt. Promising hits can become explicit seed_qualified_names for a follow-up call.
build_codemap(directory=..., seed_qualified_names=[A, B, C],
depth=1, format="mermaid", max_nodes=30)
Paste output into mermaid.live, render via npx -y @mermaid-js/mermaid-cli, or commit the .mmd to .docs/.
build_codemap(directory=..., seed_qualified_names=[...],
format="json", include_snippets=true)
Read nodes[].id (32-byte array) ↔ edges.from/to. Use qualified_name, file, line for display. snippet carries first ~5 source lines.
build_codemap(directory=..., task_prompt=..., embedding_policy="compute_missing")
First call seconds-to-minutes depending on retained-node count. Cache fills; subsequent calls are fast. Cache is shared with semantic_overlaps.
build_hypergraph(directory=..., force_rebuild=true)
build_codemap(directory=..., seed_qualified_names=[<edited-fn>], depth=2)
build_codemap automatically fires the staleness diagnostic if you forget the rebuild. compute_fingerprint flips on any .rs or Cargo.toml byte change.
| Goal | Tool |
|---|---|
| Single fn caller/callee list | rmc-call-graph (not codemap) |
| Subgraph around one fn | build_codemap(seed_qualified_names=[X], depth=1) |
| Exploratory subgraph | build_codemap(task_prompt=..., depth=3) |
| Duplicate-detection / clones | rmc-semantic-overlaps (not codemap) |
| One-symbol forensics | rmc-symbol-forensics (not codemap) |
| Diagram for docs | format="mermaid" |
| LLM ingestion | format="json" |
| Terminal-readable | format="outline", often with include_snippets=true |
| Quality over speed | embedding_policy="compute_missing" |
| Speed over quality | embedding_policy="no_rerank" (default) |
| Pattern | Invocation |
|---|---|
| Override-seed, default | build_codemap(seed_qualified_names=[X]) |
| Multi-seed cohort | build_codemap(seed_qualified_names=[A, B, C], depth=1) |
| Prompt-driven | build_codemap(task_prompt="...", depth=2) |
| Narrow (single concept) | max_nodes=15, depth=1 |
| Broad (subsystem map) | max_nodes=200, depth=3 |
| Visual | format="mermaid" |
| Annotated text | format="outline", include_snippets=true |
| All views | format="all" |
| Cached cosine rerank | embedding_policy="cached_only" |
| Compute + cache | embedding_policy="compute_missing" |
Codemap: <seeds-summary>
Snapshot: <graph_id> (<diagnostic about staleness if any>)
Nodes: <n> Edges: <m> Seeds: <s> ms: <ms>
Embeddings: <cached / computed / none>
Diagnostics:
- <line per diagnostic, or "none">
Hierarchy (top items by relevance):
* <seed> [<ItemKind>] <file>:<line>
<neighbor1> [<ItemKind>] <file>:<line>
<neighbor2> [<ItemKind>] <file>:<line>
...
Edges (sampled by weight desc): <e1>, <e2>, ...
Verdict: <leaf cluster | hub | midstream flow | exploratory>
Followups: <up-to-3 specific suggestions>
Use * prefix for seeds. Group by parent module path in the hierarchy block. For format="mermaid" outputs, return the raw flowchart LR ... block (no wrapping).
pub / pub(crate) items. Module-local private fns can't be seeded by qualified name. The diagnostic distinguishes "not indexed" from "typo" by checking whether the parent module resolves.is_callable) may not rank those identifiers' definitions; the result is dominated by code that talks about the concept. Use override seeds for pinpoint navigation.dyn ExternalTrait / generic F: Fn(..) indirect calls are inherited blind spots from Usage extraction. Local trait dispatch IS captured (resolves to the trait declaration's Item).compute_missing first-call cost: model load is ~80 MB on disk, per-node embedding is ~10 ms; subsequent calls hit the cached vector and are sub-ms per node.